GTX 970 VRAM (update: 900M GPUs not affected by 'ramgate')

What are you talking about?
Go read the PCPer link.

Its a new way they deal with memory control when they disabled the GM204 to make the 970

Ive read many 970 reviews and recent tests by 970 owners and they notice no stuttering with high VRAM usage. Other than a few reports of users being stupid and enabling every setting possible in games that doesnt use much VRAM, including MFAA and high resolution to come over 3.5GB, and complaining about FPS drop or non fluid performance, because its a faulty card, not a setting that could choke any GPU out there lol

 

I read it. But if the benchmark can only access 3.5GB VRAM, then why doesn't 980 start slowing down at the same time, since it would also be swapping from system RAM?

 

the benchmark can only access 3.5 GB in the case of the 970 because of the two separate vRAM partitions. the 980 has just one partition, thus it doesnt show the same behavior with that benchmark. at least thats how i understood it

 

EDIT: Beat by jaybee

970 = 3500 + 500
980 = 4000

Benchmark:
970; Can only access 3500MB on 970 because the 500MB is locked down by Nvidia (drivers perhaps?) and is only available on games and other software. So the benchmark start using the system RAM after hitting around 3500MB. Many apparantly use 1600MHz dual channel which have 22GB/s bandwidth. And you see 22GB/s listed in the benchmark.
980: Can access all 4000MB on the 980. No drop in bandwidth because it doesnt need to touch the system RAM. It stops around 4000MB

 

So if this CUDA benchmark is unable to access the full 4GB, how are you so sure games and other apps can?

 

lol read the PCPer link ffs.
They tested 2 games with VRAM usage over 3500MB = it accessed the last 500MB
And they had a 3% decrease in performance vs when they tested the games with less than 3500MB usage which is certainly neglible

There is a reddit link posted by n=1 where several users did the same

 

He said, she said. I've heard both sides equally.

What makes you think I haven't read the PCPer article 10 times already? It doesn't definitely prove anything.

 

And the twisting starts.
I`m out

Edit: Maybe you read the PCPer article but you certainly didnt understand it.

 

Then enlighten me with something more concrete than your assertion that the benchmark is unable to access more than 3.5GB on a 970. Because that's your whole premise up to this point, right? That the benchmark is untrustworthy?

Here, something just as concrete:

 

How on earth am I gonna prove or disapprove a video? There is literally a thousand reasons it may act this or that way

The pink flashing lights is a dead give away that the card is faulty anyway or something else is horribly wrong. You dont think the entire internet would blow up if that was the case for all 970s? Its super easy to spot unless you are blind. Come on, you know better than that

 

I'd like to know what was done to to exceed the 3.5GB limit. If it was some added post processing or something else. Perhaps it's just that part of that added feature is FUBAR. I'm not saying there isn't something odd going on, but so far it's all been circumstantial evidence without some substantial details and some structured method applied to a multitude of cards.

I also noticed the system RAM usage went well over double (from ~ 6GB to over 13GB) which is quite odd.

 

So the benchmark(software) is at fault here?

 

EXACTLY! The 970 is a faulty card!

 

Probably anti-aliasing.

It's swapping to system RAM because the 970 can't use over 3.5GB?

 

There's the problem... "Probably"...

System RAM and pagefile use also doubled...

 

Read my edit.

 

No, THAT 970 could be a faulty card

Nice try

 

So that 970 can selectively choose when to be faulty?

"I'll just act perfectly normal for the first half of this video and then act like an RMA in the second half."

BTW, first line of the video description:

So it's not a faulty GPU as the visual artifacts only appeared in the ShadowPlay recording. However, the microstutter does exist.

 

System RAM is 6GB vs 13GB and Swap file is 11GB vs 17GB. Swapping 6-7GB of extra RAM and swap file because it can't use over 3.5GB? I don't think so. I just think this example is a bad one, and likely a faulty card, not due to the 3.5GB vRAM possible issue. Bad vRAM would exhibit that failure, not slow vRAM.

 

That is obviously not typical for GTX 970. People wouldnt have needed a benchmark to run in cmd and measure bandwidth if that was the case.

Haha seriously. That was the worst example one could find.

 

What, the pink screen? That only shows up in the ShadowPlay recording per the video description. However, the microstuttering is real.

 

Nope, Im not playing this game. Its a timewaste.

I will be sitting here, enjoying my 970Ms which works perfectly without stutter, watching people waste their time posting this and that graph and this and that video and watching this blow over

Nvidia gave an official reply why the benchmark got a drop in bandwidth and why its a flawed test, the card have been proven many times to go over 3.5GB without any performance drop,ive yet to see a review that says it stutters (many reviews measure that now) or have a minimum FPS far below 980, or that its performance drop tremendously once over 3.5GB usage.

The card doesnt have any issues, people didnt have any issues with it before until this benchmark came along either.

The Internet

 

Yeah, who had the brilliant idea to post a desktop 970 problem in a laptop forum anyway?

 

The person who knows that the 980m/970m are variants of the desktop 970, so it seems obvious to pose the question of whether the desktop 970's problem extends to these mobile cards.

 

^Wow, this guy. Sarcasm detector broken.

 

...it hasn't had it's coffee yet.


Did anyone see Nvidia's response?

NVIDIA Responds to GTX 970 3.5GB Memory Issue | PC Perspective

 

^are you serious? Did you read through this thread, and in particular post #42?

Anyway, after a reading a couple forums the impression I get is that the problem only occurs when a game requires >3.5GB of vram. Emphasis on REQUIRED because allocated amount =/= actively used amount, which explains why Watch Dogs at 1.78x DSR (1440p downsampled to 1080p) which can suck up 3.8GB of vram did not incur any performance penalties.

My take is that this isn't likely to be a problem at 1080p, or even 4K on a single GPU (you'll run out of core power long before the vram issue rears its ugly head). 4K on dual or tri-SLI setups is where I can see this being a potential problem.

 

Well, obviously..... no. I think I jumped to page 6 ...or 7 to start.


...but that's why I asked if it was shown.


...for just that possibility.


EDIT:

That's now, that's not later on. You get a GPU with 8GB to "future-proof" yourself against future game requirement for the same reasons you choose a DX12 card before DX12 arrives. Having games want more than 3.5GB of VRAM is highly probable over the 980m's lifetime and it would be unforgivable if this bug was the limiting factor in the future.

 

It seems my 980m 8GB does have the bug, but I'm not sure I should care a lot. If the card is not going to have an issue until I hit 7.5GB, big whoop. I don't understand why the end of the first 4GB is fine when the last 4GB has the issue. I sort of expected the issue to occur at 3.5GB and 7.5GB.

...ooops! ...sorry they're out of order.

...and for the back2back posts.

 

GL hitting 8gb of vram without getting 3fps in the game

remember 3% performance loss of 0fps = 0 fps

 

Thanks for sharing. But does this mean the 8GB also has 500MB of separate vRAM? It exhibits the same behavior as the 4GB 970, which I find quite odd. I think the benchmark is simply flawed and irrelevant.

 

Looking it up, thegreatsquare's notebook doesn't seem to have the 980M in a headless configuration (it is used as the primary display adaptor with the Windows desktop taking up some VRAM on the card) so that may have caused those results, based on this post.

 

Right, so that benchmark is useless really for determining this so-called issue.

Not only that, I would suspect any storage medium that was near capacity would result in some performance issues. If the game really *needs* 3.8GB then sure it will stutter a bit as it's thrashing the data in and out of system RAM or worse yet pagefile. I think this is all much to do about nothing.

 

Dont think my P650 SG has any issues at all...

Stock


190/380


200/380

 

Anandtech article:

http://www.anandtech.com/show/8931/nvidia-publishes-statement-on-geforce-gtx-970-memory-allocation

Sent from my Nexus 5

 

You are using a GT72 which has MUX which means your display while the 980M is running is on dGPU only which means windows is taking up your vRAM as well, which is why the last few runs are being used up. Just like with my screenshot on the first page.

Just need a game with seriously high resolution textures to show up and you can. vRAM usage != core drain. You can have super high resolution, super low-quality textures. They'll be sharp as hell but look super terrible.

See above, Wingnut. The bench was designed to work when someone was plugged into the iGPU for their only monitor so that the system doesn't use the vRAM on the card itself at all.

Because you have an iGPU installed. When you get your P750ZM you'll find that it "does" have the problem, actually.

 

So can anyone with 900M and Windows 7 run the benchmark with the compositor disabled? Since it seems like nobody can run in headless mode on notebooks.

 

Some food for thought?

But yeah, I'd like to see some FCAT numbers as well, that would be the most telling.

 

I would agree, but Nvidia then came out with some statement basically admitting to ghetto-rigging the desktop 970 to some extent so there has to be more to it.

 

Ahhhmmm...Kepler?

880M test:

 

Exactly. Its flawed because it can`t access the 500MB partition which games and other software do. Which I guess the driver control.
Which is why reviews that tested games and GPGPU software never saw anything fishy

 

Again, you're on a non-optimus enabled machine with Windows 8.1 (vRAM hog). You would get those results.

 

You have to run in headless mode or disable DWM for accurate results. That drop is from Windows taking up a few hundred megs of VRAM.

 

Ok Ok...don't kill me.

 

Ok, no problem...

 

I just rebooted into intel GPU mode Fn+F5 (I/D GFX) and tried to run the test. Program crashes upon loading.

 

More detailed article from Anandtech.

Looks like the GTX970 actually has 3.5GB of fast (196GB/s) vram and .5GB of slow (28GB/s) vram that can't be read simultaneously, as originally reflected by the benchmark. Mystery solved. Doesn't seem like there's any significant performance hits, but that's still TBD.

 

Anyone tried this with 980M 8GB? Does it have similar allocation of fast vs slow VRAM?

 

the desktop 970 is the only 9xx series card with more than one memory partition, so rest easy

 

So... 970 is a jury rigged card. Nice. I don't understand why they even ELECTED to make such a card via design. That's probably the reason why it was so cheap, come to think of it.